Permutation Coding Technique for Image Recognition System - Neural Networks and Micromechanics - page 48

RLDs are similar to Rosenblatt's descriptors but are applied to the local area of
the image. In the first version of our recognition system, each random detector was
applied to its own local area, selected randomly in the image. This recognition
system was named LIRA. LIRA was tested on the MNIST database and showed
sufficiently good results - 55 errors (see Table 3.1). One of the LIRA drawbacks is
the sensitivity to image displacement. We compensated this with distortions of
input images during the training process; however, this method cannot be used for
large displacements. For this reason, we developed a more sophisticated image
recognition system.

4.3 General Purpose Image Recognition System Description

The scheme of the general-purpose image recognition system is shown in Fig. 4.1 .
The base of this system is a multilayer neural network. The first layer, S (sensor
layer), corresponds to the input image. The second layer, D 1 , contains RLDs of the
lowest level, while the layer D 2 contains RLDs of the highest level. The associative
layer A contains associative elements, which could be represented by groups of
neurons, and the layer R contains the output neurons. Each of these neurons
corresponds to the image class under recognition. The scheme of the lowest-level
RLD is presented in Fig. 4.2 .

D 1 D 2

d 11

d 21

A

a 1

R

S

d 12 d 22

a 2

a N

d 1M

d 2M

Fig. 4.1 Structure of a
general-purpose image
recognition system